Toward the Explainable Soft Prompts: How does Prompt-tuning Exploit a Multilingual Pre-trained Language Model?

Anonymous

Abstract: Since training soft prompts is a parameter-efficient way to tune a Pre-trained Language Model (PLM) on a target task, recent works suggest various training methods utilizing soft prompts. However, few studies investigate the explainability of soft prompts, which is critical to enhancing the confidence of PLM in a real-world scenario. To deal with the problems of the unexplainable soft prompts, this study explores the effects of Prompt-tuning v1 (Lester et al., 2021) and Prompt-tuning v2 (Liu et al., 2022) on PLM. More precisely, we conducted the experiments using a multilingual GPT to generalize our observations not only on tasks but also on languages. We first confirmed whether soft prompts are gathered according to tasks or languages, and then analyzed how soft prompts utilize PLM in terms of the two main architectures of GPT: the attention mechanism and the activated neurons. As a result, we conclude that soft prompts are trained while employing the knowledge PLM obtained during pre-training to solve the target task, which is consistent with language. Our findings reveal that deep soft prompts are explainable when they can employ varying knowledge from every layer.

Paper Type: long

Research Area: Interpretability and Analysis of Models for NLP